Reproduce “Cyclotron: A Recurrence Language for Expressing Interprocessor Communication” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shiv Sundram, Akhilesh Balasingam, Nathan Zhang, Kunle Olukotun, Fredrik Kjolstad
Cyclotron lets programmers describe communication routes as recurrence relationships between points in an iteration space, separately from mapping dimensions to processors or time. Its compiler translates them into distributed computation and communication. For blocked Cholesky on 16 Perlmutter CPU nodes, the best schedule ran 1.68× the ScaLAPACK speed at one tested configuration; other schedules were slower, and winners varied.
Key insights
- The same algorithmic recurrence can support different communication patterns and space–time mappings, enabling systematic exploration without rewriting the computation.
- Communication strategies are hardware- and workload-dependent: for example, broadcast performed poorly on one tested cluster but was reported best on a separate densely connected system.
- Dependency-driven scheduling can expose overlap across operations and kernels without requiring a global barrier, in addition to managing communication correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.